| has_test_audio: true # set this to False if you don't have access to the test audio (i.e. training and validation only) | |
| has_test_labels: true # set this to False if you don't have access to the test labels | |
| wav_crop_len: 5 # Length of cropped files in seconds | |
| data_path_base: ../data/production_data | |
| n_classes: 66 # Number of classes | |
| pretrained: true # Use pretrained model | |
| backbone: tf_efficientnetv2_s.in21k # image classification model (from list_models) | |
| in_chans: 1 | |
| num_workers: 4 # Number of parallelized CPUs | |
| include_val: true # Validation-set included / excluded | |
| max_amp: false # Experimental feature | |
| # Training Hyperparameters | |
| n_epochs: 18 # Number of epochs | |
| lr: 0.0017 # Learning rate | |
| weight_decay: 1.0e-05 # Weight decay | |
| label_smoothing: 0.1 # Label smoothing | |
| batch_size: 32 # Batch size | |
| sample_rate: 44100 # Sample rate | |
| # Mel Spectrogram Hyperparameters | |
| # see Torchaudio Documentation to understand these | |
| n_mels: 128 | |
| n_fft: 2048 | |
| fmin: 400 | |
| fmax: 22000 | |
| power: 2 | |
| top_db: 80.0 | |
| win_length: 2048 | |
| hop_length: 1024 | |
| # Normalization | |
| mel_normalized: true # Mel normalization as documented in Torchaudio (normalized=True) | |
| minmax_norm: false # Apply minmax normalization on spectrograms | |
| # Augmentation Parameters | |
| impulse_prob: 0.15 # Impulse probability | |
| noise_prob: 0.15 # Noise probability | |
| max_noise: 0.04 # Noiseinjection amplitude | |
| min_snr: 5 # signal-noise ratio (Gaussian & Pink Noise) | |
| max_snr: 20 | |
| mixup: false # Apply mixup augmentation | |
| specaug: false # Apply OneOf(MaskFrequency, MaskTime) | |
| specaug_prob: 0.25 # Probability to apply spectrogram augmentation | |
| mixup_prob: 1 # Parameter of a symmetric Beta distribution, 1=uniform distribution | |